Volume 20, Issue 12, Pages 2784-2791 (September 2017) Dynamic Rearrangement of Cell States Detected by Systematic Screening of Sequential Anticancer Treatments  Simon Koplev, James Longden, Jesper Ferkinghoff-Borg, Mathias Blicher Bjerregård, Thomas R. Cox, Janine T. Erler, Jesper T. Pedersen, Franziska Voellmy, Morten O.A. Sommer, Rune Linding  Cell Reports  Volume 20, Issue 12, Pages 2784-2791 (September 2017) DOI: 10.1016/j.celrep.2017.08.095 Copyright © 2017 The Author(s) Terms and Conditions

Cell Reports 2017 20, 2784-2791DOI: (10.1016/j.celrep.2017.08.095) Copyright © 2017 The Author(s) Terms and Conditions

Figure 1 Systematic Screening of Sequential Combinations of Anticancer Drugs to Identify Temporal Synergy (A) Schematic illustration of temporal synergy where changes in cell number can be induced by sequential treatment with drug α followed by drug β. Sequentially effective combinations could be time-dependent or reflect the dynamics of classical simultaneous synergy. (B) Cytotoxicity measured by high-content imaging and quantified using a global synergy model, which prioritized additional validation experiments for 200 drug combinations. Cells, in 384-well plates, were treated with drug α for 24 hr and then treated with drug β for 24 hr at 4 doses. Common mechanisms explaining sequential synergy and antagonism across multiple drugs were then investigated. (C) Experimental conditions for systematic screening of sequential combinations between 100 drugs, including timing and concentration series for 3 distinct types of experiments: drug α alone, in 8-point dose response, where cells were assayed after 24 hr; drug α0, where cells were treated with drug for 24 hr at 1 dose, and then the drug was removed and cells were assayed after 48 hr; and drug αβ, where cells were treated with 1 dose of drug α for 24 hr and then drug β in 4-point dose response (5-point dose response in the validation screen). All experiments were performed in triplicate, generating a total of ∼250,000 data points. Cell Reports 2017 20, 2784-2791DOI: (10.1016/j.celrep.2017.08.095) Copyright © 2017 The Author(s) Terms and Conditions

Figure 2 Global Bayesian Model of Cell Viability Data (A) Global Bayesian model of cell viability data. In total, the model consisted of 45,000 parameters, over which posterior probability distributions were fitted using a Metropolis-Hastings algorithm, assuming sigmoidal dose-response curves and Bliss independence between consecutive treatments. Prior distributions over all parameters were assumed. Each observation of type α, α0, and αβ carried equal weight in the Bayesian inference. (B) Special selector variables (λ) were used to enable the αβ data to influence the baseline fit for a more conservative estimate of sequential effects. Synergy was quantified as the difference between the expected baseline and the model fit, with antagonism associated with negative values of this measure. (C) Example of validated model fit in A375. Left: conservative fit of baseline dose-response curve for lomustine based on average posterior parameters from supporting data points, including controls (experiment types β and β0) in addition to all combinatorial experiments (αβ) that involved lomustine. To illustrate their influence on the baseline fit, each point was scaled by effects from residuals and non-lomustine drugs according to the 3-factor Bliss independence model. Right: average fitted dose-response curve for lomustine pretreated with amifostine, showing a synergistic sequential effect, p < 0.0005. (D) Distribution of the posterior MCMC frequencies of selector variables for all 10,000 sequential combinations estimating the likelihood of drug interaction and whether the interaction was synergistic or antagonistic. These “λ scores” were multiplied by −1 for antagonistic combinations yielding a range of [−1, 1], where −1 corresponds to the most antagonistic combination and +1 corresponds to the most synergistic combination. Cell Reports 2017 20, 2784-2791DOI: (10.1016/j.celrep.2017.08.095) Copyright © 2017 The Author(s) Terms and Conditions

Figure 3 Sequential Synergism and Antagonism among Anticancer Drugs Are Common in A375 and PANC1 Cell Lines (A) Heatmaps of average posterior synergy measures (difference from baseline) for all 10,000 sequential combinations tested in PANC1 and A375, where blue indicates strong antagonism and red indicates strong synergy. Rows and columns correspond to first and second drugs, are arranged by hierarchical clustering, and are colored by classes of drug mechanisms. (B) Average synergy measure by drug mechanism showing, in both cell lines, increased synergy following secondary treatment with alkylating agents and strong antagonism following secondary treatment with tubulin modulators. Significance was assessed by permutation tests. Cell Reports 2017 20, 2784-2791DOI: (10.1016/j.celrep.2017.08.095) Copyright © 2017 The Author(s) Terms and Conditions

Figure 4 Regression-Based Interpretation of Schedule-Dependent Synergy in Terms of Drug Mechanisms, Protein Targets, and Associated Molecular Pathways (A) Cross-validation error of classes of regression models for the first (α) and second (β) drugs, illustrating that drug protein targets and/or pathway activity did not significantly improve predictive power over mechanism alone. However, some particular meta-features, such as the protein target of the pretreatment combined with the mechanism of the secondary treatment, did increase predictive power. The line represents the greedy conjunction of the best performing models added one at a time in the order of their individual cross-validation performance. All fits were controlled for overfitting by using the hyperparameter value yielding the lowest 10-fold cross-validation error. (B) Mean synergy measures quantified by repression coefficients illustrating synergistic and antagonistic effects from individual drugs and drugs grouped according to their described mechanism, as either the first (α) or the second (β) treatment. Cell Reports 2017 20, 2784-2791DOI: (10.1016/j.celrep.2017.08.095) Copyright © 2017 The Author(s) Terms and Conditions